Add dataset card

README.md

@@ -1,3 +1,79 @@
 ---
 license: mit
+task_categories:
+  - image-segmentation
+  - object-detection
+language:
+  - ar
+  - fr
+tags:
+  - coco
+  - instance-segmentation
+  - medicine
+  - medication-boxes
+  - mask-rcnn
+pretty_name: Medication Boxes (Arabic / Latin)
+size_categories:
+  - n<1K
 ---
+
+# Medication Boxes — Arabic / Latin
+
+This dataset contains photos of medication boxes, with packaging text in Arabic and Latin script (French and others). It is annotated for **COCO instance segmentation** and was built for **MediSeG**, which segments each box so the text on it can be read afterwards.
+
+- **One class:** `1 = medicine_box` (`0 = background`)
+- **Format:** COCO JSON (polygons + bbox)
+- **Images:** original resolution, never resized
+
+## Structure
+
+```
+train/        540 images
+val/           68 images
+test/          68 images
+annotations/
+  instances_train.json
+  instances_val.json
+  instances_test.json
+```
+
+| Split | Images | Instances |
+|---|---:|---:|
+| train | 540 | 806 |
+| val | 68 | 121 |
+| test | 68 | 100 |
+| **Total** | **676** | **1027** |
+
+Each `images` entry also has three extra fields: `source`, `original_split` and `original_file`.
+
+## Sources
+
+| Source | Images | Instances | Original annotation |
+|---|---:|---:|---|
+| `main_ar_fr` | 557 | 874 | YOLO-seg, 4-point polygons |
+| `medicine_packv2` | 119 | 153 | COCO polygons (24 drug classes merged into `medicine_box`) |
+
+The `drugs` source (1,068 images) was **left out** because it has bounding boxes only and no segmentation.
+
+## Preparation
+
+- 80/10/10 split (seed 42), **grouped by source photo**. Near-duplicates (dHash of the object crop) go into the same split, which fixes the train/val/test leakage in the original splits (128 images affected).
+- Checks: no corrupted images, no invalid annotations, no empty masks, no exact duplicates (md5).
+- Validation: every split loads with `torchvision` `CocoDetection` + `wrap_dataset_for_transforms_v2`, and a `maskrcnn_resnet50_fpn(num_classes=2)` forward pass gives finite losses.
+
+## Usage
+
+```python
+from huggingface_hub import snapshot_download
+from torchvision.datasets import CocoDetection, wrap_dataset_for_transforms_v2
+
+root = snapshot_download("ApyHTML19/Medication_Boxes_AR_Latin", repo_type="dataset")
+ds = CocoDetection(f"{root}/train", f"{root}/annotations/instances_train.json")
+ds = wrap_dataset_for_transforms_v2(ds, target_keys=("boxes", "labels", "masks"))
+img, target = ds[0]
+```
+
+## Limitations
+
+- Small dataset (676 images), a single class, no per-drug labels.
+- The `main_ar_fr` polygons have 4 points, so masks are quadrilaterals that approximate the box outline.